Sourcing and Vetting Alternative Data

A vendor pitches a dataset of credit card transactions covering thousands of merchants and promises it predicts retail earnings weeks before they're announced. The pitch deck has a beautiful backtest. Before a single dollar is spent, a research desk needs to answer a much less glamorous set of questions: how was this data collected, who else already has it, is it legal to trade on, and does the sample actually look like the world it claims to describe. Sourcing alternative data is mostly this diligence work; the modeling comes after.

Where datasets come from

Vendors fall into a few recognisable categories, and each carries its own risk profile. Some aggregate data that already exists for another purpose — a payments processor selling anonymised transaction data, an email receipt aggregator selling parsed purchase confirmations. Others generate data specifically to sell to traders — satellite tasking companies, web scrapers, survey panels. A third group resells or blends other vendors' feeds, adding a layer of opacity about the original source. Knowing which category a vendor falls into changes what questions matter: a reseller's biggest risk is a licensing chain you can't fully audit; a data generator's biggest risk is a sample so small or so biased it doesn't represent the thing it claims to measure.

The diligence checklist

Before committing budget, a researcher works through a standard set of questions, roughly in this order of importance:

Question	Why it matters
How is the panel constructed and how large is it?	A small or self-selected panel (say, users of one budgeting app) may not represent the population it claims to proxy for.
Is the history point-in-time, or was it reconstructed later?	Data that was cleaned or expanded with hindsight overstates how good a live backtest would have looked.
Does panel composition change over time?	A panel that grows from 10,000 to 500,000 users mid-sample makes early and late periods barely comparable.
Is this data legal to trade on?	Some feeds sit close to material non-public information or violate a data subject's terms of service; see below.
Who else already licenses this feed?	A signal that a dozen multi-strategy funds already trade is close to arbitraged away before you start.
What's the true cost including onboarding?	List price is rarely the full cost — cleaning, mapping to tickers, and legal review all add real time and money.

Worked example: evaluating a credit card panel

A vendor offers a panel claiming to cover 2% of US consumer card spend, with three years of history, priced at $300,000 a year. Diligence turns up: the panel's user base doubled in year two after a partnership deal, meaning the early sample is thinner and noisier than the later sample looks; the vendor cannot say what fraction of transactions are business versus personal, which matters for retail-sales proxies; and a call to the vendor's sales team confirms two competing hedge funds already hold licenses. None of these facts alone kills the deal — panel growth can sometimes be adjusted for, business-mix ambiguity can be tested against known quarters, and being third to a feed still leaves room for a distinct signal. Together, they change the negotiation: a research team armed with this diligence asks for a lower price reflecting the panel's early-period weakness, requests raw transaction-level detail rather than pre-aggregated series so it can build its own point-in-time cuts, and budgets extra research time to find an angle the two existing clients haven't already traded.

In practice

• Ask for a trial period with historical data, not just a live feed. A vendor confident in their product will let you backtest before committing to a multi-year contract.
• Talk to the vendor's data-collection team, not just sales. Sales will describe the panel as it is meant to work; the collection team knows where it actually breaks down.
• Budget real time for mapping. Raw alt data almost never arrives ticker-ready — see Brand, Merchant and Subsidiary Mapping for the specific problem of turning merchant or brand names into tradeable securities.
• Treat legal review as a gate, not a formality. A dataset that turns out to embed material non-public information can taint an entire desk's other positions; see MNPI and Data Licensing Risk.
• Revisit the diligence annually. Panels change composition, vendors get acquired, and a feed that was clean and exclusive at signing can quietly become neither.